Efficient Anomaly Detection Using Self-Supervised Multi-Cue Tasks.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep anomaly detection methods leverage self-supervised learning (SSL) and geometric transformations.
• Existing SSL methods for anomaly detection struggle with fine-grained features and exhibit type-dependent performance.

Purpose of the Study:

• To develop advanced discriminative and generative tasks for enhanced deep anomaly detection.
• To improve the detection of fine-grained, object, style, and local anomalies.

Main Methods:

• Introduced three novel SSL tasks: piece-wise jigsaw puzzle (structure), tint rotation recognition (colorimetry), and partial re-colorization (texture).
• Incorporated an attention mechanism for object-oriented re-colorization using contextual border color information.
• Developed a stable out-of-distribution detection function and experimented with score fusion techniques.

Main Results:

• The proposed method significantly outperforms state-of-the-art approaches.
• Achieved up to 36% relative error improvement on object anomalies.
• Demonstrated up to 40% relative error improvement on face anti-spoofing tasks.

Conclusions:

• The novel SSL tasks effectively capture finer features crucial for anomaly detection.
• The proposed method offers superior performance and stability across diverse anomaly detection challenges.
• This work advances the capabilities of deep anomaly detection, particularly for fine-grained and complex scenarios.